I have an idea that enables us to make the world a better place,
Like all really good ideas, this idea begins with a journey along the way.
The key figures during this trip are me and a professor at the University of Chicago,
Which examines crime and the criminal justice system,
And one of my friends, a professor at the Ivy League Medical School.
And so we were, at least two prominent academics
Money heading for Route 95.
(Laughter)
We had to spend a few hours on our way from New York City to New England,
I tried then to try to use these talking skills
Which is famous for its professors.
I paid attention to my friend and told him,
“Tell me the biggest mistake you’ve ever done in your whole life.”
He stopped a little and then turned to me and said,
“I have an idea, why do not you start first?”
(Laughter)
I said to my friend
“I asked the question, why do not you start first?”
My friend told me about a time when he was in the emergency department
And entered a patient complaining of chest pain.
The protocol is customary in such cases
Is to do a check for heart enzymes
To know the contents of the blood in order to be sure
Whether the patient is suffering from a heart attack at the time.
Enter the patient and receive the examination.
The level of enzymes was above normal,
The patient is usually held in intensive care.
But before doing so, my friend enters the waiting room to see the patient,
To see him personally at first.
The patient was sitting there and eating watermelon as a snack.
My friend spoke to him for a few minutes,
Then go back to meet the rest of the team.
And now the rest of the team –
Doctors and nurses on emergency task –
They did not see the patient.
All they have seen is the data in the table and the result of the examination is above normal.
They said, “We have to go and take the man into the intensive care unit”
My friend replies, “No, no, no.
I went and met the patient, he was very satisfied and very warm.
He is sitting there and having a snack.
I think it’s okay, let’s just leave it alone. “
Half an hour later,
The man enters into a heart attack and has to be rushed to the operating room.
This is an illustration of the lesson we have learned
On the behavioral economy and psychology it, how can the human mind simply
To be misled by information that is irrelevant, but very important.
This made me think of a problem that was my research center
Who is co-director of the crime lab at the University of Chicago
Works to solve it for many years,
The problem of the prison system in the United States.
Millions of times every year
Judges must decide whether or not someone will be arrested,
Where that person awaits trial:
Can they go home or should they stay in prison?
The decision is supposed to be the law
To rely on judges’ expectations
What can be done by the defendant if it is released frankly:
Is that person vulnerable to escape from the country?
Is that person a threat to public security?
This is a very fateful decision.
If the judge decides to put you in jail, you will stay there for a while
Month or three months on average, and sometimes for much longer than that.
The corresponding decision is as follows:
If the judge decides to release someone, who is likely to commit new crimes.
Which would be very terrible in its consequences.
This is a very difficult decision for the judge
For the same reason that made the emergency room decision
A tough decision for my friend the doctor.
Emergency doctors have at least an advantage, such as heart enzymes
Which helps them to take these kinds of decisions.
Judges are given a package of Manila files by some information
The reason for the arrest of that person
And the previous criminal record,
The judges must make a decision within their heads.
Let us imagine the madness of this
It must be borne in mind that the same judge who spends the whole day reads the transcript between folders
And makes decisions that may change the course of life and fate of some people
They then return home and want to relax a little at the end of the day
By watching a movie on TV
And to help make this critical decision
The judge has access to the Netflix platform
Which use some of the most complex automated learning techniques
On the face of the earth
Which can predict any film the judge will like.
Why not use some of these techniques
Which have been applied very effectively in the commercial sector
To help us solve these policy problems is also very important?
Now, in order to know whether this will really be useful or not,
I will use it to be better for beginners to be more knowledgeable
What is automated learning and how it works.
Let me briefly explain to you
A standard problem in computer science, called the analysis of conscience.
Here’s what I mean:
It is a section of text
And try to determine the author’s passion at the time:
Is the author trying to convey a positive or negative emotion?
If I show you how it looks
For a consumer product selected more or less random
A banana cutter (Hatzler 571)
(Laughter)
Now we have a critical review of the product by (Rivette E)
“I bought these so I could cut the bananas for breakfast faster.
The time I spent with this attempt was spent cleaning this tool. “
(Laughter)
“Cleaning it is not easy.
You have to stick between each blade to clean it well. “
Now, when we read this, we can easily and simply consider this a negative assessment.
We can confirm our estimate of the order given the number of stars in the rating
Two of only five stars.
Here’s another review by Uncle (Bucky), he says
“Great gift”
(Laughter)
“When I discovered that I had to peel the banana before using the tool”
(Laughter)
“You are working much better.”
Five-star rating.
These are added by (QTip)
“It’s confusing, there’s no way to tell if this is a regular or standard banana cutter.”
(Laughter)
“Marking them would be very useful.”
This is another comment by J. Anderson:
“The angle is wrong.
I have tried this banana cutter and found it completely unacceptable.
As shown in the picture, the cutter is curved from left to right
I have all the bananas in the other direction. “
(Laughter)
Now after reading these comments and criticisms
You will discover that it is very, very easy for us to do so
This gave the first computer scientists
How to make a computer do it too.
Why do not we delve into and examine ourselves only by how we do this,
And then we try to program the computer to do as we do exactly?
Here we have the results of a study trying to analyze emotions
Using so-called programmatic progress to review movies.
Our film review data collection
We find that half of them are positive reviews, and the other negative reviews.
Thus, accuracy of 50% would be completely random.
You become a group of programmers
And who somehow sort out the words you expect to see
In positive reviews, and negative reviews.
Here we have some positive words
Which you think you can see in good reviews
And some words that can be seen in negative reviews.
When you do so, you have 60% accuracy.
Now this is better than random guessing,
But not much better.
These are the challenges that computer scientists have faced in this area.
Even if the problems are simple
It has turned out to be very, very difficult
Programming the computer to be able to do what we do
And get good performance.
The reason for this is that it turned out to be much harder
Is the introspection and discovery of what we actually do when we do these tasks fully.
Psychologists called my friends on so-called “illusionists”
No progress has been made in this area until
It turned out to computer scientists that we just had to forget completely
We knew how to do these things ourselves
And to change these tasks to just drastic training on data.
Here’s how this will look in analyzing film reviews:
You will come with a great sample of movie reviews
Which will determine whether they are positive or negative
By Star Rating
The computer will learn
Which of the words often appears in positive reviews
Any words appear in negative reviews.
OK?
Then use these words as algorithms you predict
Future Reviews.
Once you have reached that progress with data,
These are the words that the computer learns,
Which the machine learns,
An indicator of positive and negative reviews.
Now we can get the accuracy of this system by 95%
This is what I think is the magic behind learning the machine really.
You can see how we can apply it
Such as pre-trial releases.
Let the computer learn any of the attributes and information of the case,
Or a combination of features and case information,
Are the most likely to see the escape from the country or threaten public security.
I was just working as part of a research team a few years ago
We try to build predictive algorithms for pre-trial release
To see if that is useful to judges.
We were doing this using data from a huge and unknown American city
With 8.5 million people.
(Laughter)
What we discovered is
It’s not really hard to build algorithms.
You can download a free software application online
And discover how to do this.
The hard part here is the experience of algorithms
And whether it will make the world truly better.
For Netflix, this is not difficult.
Everything you do (Netflix)
It is a kind of private content online.
But the experience of algorithms in the real world surrounding policy applications
It is often complicated.
This is an issue that will only be solved with difficulty,
Because of the inability to fabricate a random trial,
It is a difficult issue for sociology,
Not the difficult problem we face
Computer Science.
It is very difficult
That group of people who think
In taking these machine learning tools
To the public policy arena
By passing the pilot phase of the project
And take tools directly from the planning stage on the computer to the real world.
This I think is a big mistake.
It is possible to build a tool without awareness or caution
It turns upside down and ends up making the world a worse place and not a better place.
For the project we are working on so far
The hardest part is knowing how the tool is being tested
And make sure they will be really helpful.
The way we designed the tool was built on two visions.
Please note why this issue is difficult in a case before the trial.
We are creating an algorithm base
Which states that we give priority to those most vulnerable to risk-causing
To be imprisoned and release the rest.
This algorithm will be used by all expectations to free someone
The judge has locked him up.
When you want the algorithm to do this,
We can not know what that person would have done if he had been released
Because the judge has already imprisoned him.
That’s why we will have this difficult data loss problem.
In contrast if
If the algorithm wanted to imprison someone who had been released by the judge,
We will not have a problem evaluating
Because we know the effect of putting someone in jail
On his offer to flee the country or threaten public security.
The fact that this person is in prison erases the risk of being put on trial
Or re-arrest.
This is vision number one
It is that the challenge of missing data is a one-sided weapon.
The second vision that helps us here
Is that in the big city where we were doing our job
That the cases are assigned to the judges more or less random.
What this means is that we have a sample of judges
Who listen to similar volumes of issues.
It turns out that judges differ greatly
For their perseverance and tolerance.
If you have what we will do in this case.
Imagine that we have two judges:
One is a lenient judge who releases people in 90% of cases,
The second is more assertive and people are released in 80% of cases.
We can simply compare judges’ performance
When they become more assertive
And how to choose the algorithm to be more assertive
Which would be a fragile experience to perform the algorithm.
Before you, how it will look.
Here you have the lenient judge who releases people in 90% of cases.
We can see all the events of the judge’s release of people.
The algorithm will tell us if we want to be more assertive
We move from 90% release to 80%
The algorithm will only tell us:
To verify the highest probability of danger in 10% of people in the size of cases of the judge
We prioritize them for imprisonment.
Now the release rate has dropped to 80%
We can see the proportion of crime that we would have been able to obtain,
And then we can compare this to how the judge
Down to release 80% instead of 90%.
This gives us a way to compare poorly
Between the performance of the algorithm and the performance of the judge
In terms of a range of comparable cases,
Focusing on the algorithm task
Where we do not face the problem of missing data,
The algorithm only selects people to lock them
Of the large number of persons released by the judges.
Now, after solving the question of appreciation
The test problem,
We can do some policy simulations to suggest what will happen
If we actually follow the algorithm rule
Rather than the usual practices in the criminal justice system.
What we found is that if you follow the algorithm’s recommendations
You will be able to reduce crime by 25%
Without the need to place one additional person in prison.
Alternately, you can cut prison congestion by 42%
Without increasing the crime rate at all.
And the reason that the algorithm is capable of achieving much greater gains
On the current status of the criminal justice system
Is that, according to what we see in the statements, the judges,
Just like an emergency doctor my friend,
They are scattered with information that is not relevant but very important
On these issues.
This is especially true in the most likely cases of danger
In the respondent community.
What I have just done is that I have seen you
The positive side of applying machine learning technology to these policy issues.
There is also a potentially negative side,
The probability that these algorithms,
Once we apply them to policy issues,
Perhaps the issues of criminal justice in particular,
It is likely that we will gain some results
But you will settle other things we care about, such as justice.
You can see why people worry about this.
In the city where we operate,
89% of people in prisons are a minority –
In a city with a total population
Not in one way or another 89% minority.
People are keen to use machine learning
To solve these issues, I think they are right somehow we discovered it
It is if you create an algorithm
At the base of the release completely ignore this problem.
It is certainly possible to build a tool that will make the problem,
If any, a little worse.
But what we also discovered was that if you created an algorithm
Draws its attention to this issue,
You will be able to design support for the decision
At the same time, it can reduce the crime rate
Reducing prison overcrowding and racial disparities
In the criminal justice system as well.
How does the algorithm help you do this?
Well, what element, however,
But it is just irrelevant but very important in court?
What is implicit bias
But it is just a kind of introspection and self-reading?
The algorithm is not susceptible to these challenges
Which face human judgment and decision-making.
I think it’s exciting to order bail out
Is that it is just one statement
From a larger policy issue
Where it depends on the expectation of a human being,
But can be communicated initially through machine learning algorithms.
There is an existing controversy
About whether it’s a good idea or a bad idea to move on with these algorithms
From the commercial to the public policy.
Do I do it or not?
I think this is a really wrong style
In portraying the controversy and formulating the question,
Here is an exercise to think about why.
Imagine that I can magically bring you back in time
To the beginning of the twentieth century.
You will come there and talk to people about this new technology
Which were in sight
Which will quickly become one of the main causes of death
And have a significant adverse effect on the surrounding environment.
So now I think a little of us here relatively
They would argue that we did not have to invent cars with internal combustion engines.
Imagine how life would be without cars.
It was not to have something like this economic growth
Which we have witnessed over the past 100 years.
Our lives would become depleted in countless ways,
We would not have to make trips on the roads.
(Laughter)
So I think the debate is right
Which should be about using machine learning
In policy applications over the next 10 years
Is not whether these new techniques will be taken,
but how.
Thank you very much.
(clap)
”
Be First to Comment